Abdel Aziz Taha, Alexandros Bampoulidis, Mihai Lupu (2019)

Machine learning research, e.g. genomics research, is often based on sparse datasets that have very large numbers of features, but small samples sizes. Such configuration promotes the influence of chance on the learning process as well as on the evaluation. Prior research underlined the problem of generalization of models obtained based on such data. In this paper, we deeply investigate the influence of chance on classification and regression. We empirically show how considerable the influence of chance such datasets is. This brings the conclusions drawn based on them into question. We relate the observations of chance correlation to the problem of method generalization. Finally, we provide a discussion of chance correlation and guidelines that mitigate the influence of chance.

Read the full paper here